Could these quantized models make MTP (Multi-Token Prediction) faster when used in conjunction with larger Gemma 4 models?
alt Hacker News